From an application point of view:

	Objective may not be defined if certain constraints are violated. For example, the steady-state errors of a dynamical system are undefined if the system is not stable. Important for real-time applications.
	May have to terminate the optimization process after a prescribed amount of time, in which case it may be crucial that the sub-optimal solution at least satisfy some hard constraints.
	In the context of optimal design, tradeoff exploration cannot meaningfully take place if some hard constraints are not first satisfied. It is thus of great interest to produce iterates that all satisfy these hard constraints.

From an algorithmic point of view: 

The line search criterion can be based on the decrease of the objective function, i.e., there is no need for an artifical "merit function".